MSCE turns LLM agent memory into reusable skills, boosting long-horizon tasks
New training-free framework lets AI agents evolve skills from experience without retraining.
A new paper from multiple Chinese universities and collaborators introduces MSCE (Memory-Skill Co-Evolution), a training-free framework that upgrades how LLM agents handle long-horizon tasks. Instead of treating past experiences as passive context, MSCE organizes agent history into grounded step traces, reusable procedural policies, and declarative environmental cognition. It crystallizes high-confidence policies into callable skills that retain evidence links, applicability boundaries, decision guidance, verification rules, and reliability estimates.
MSCE also introduces reflection-weighted value backfilling, which propagates sparse terminal rewards through dense local self-reflections to produce calibrated trace values. This governs memory and skill evolution without additional training. Experiments on EvoAgentBench and LoCoMo show MSCE significantly outperforming existing skill-augmented and memory-driven agent baselines, with strong cross-domain transferability and lifelong evolution capabilities. The work is submitted to EMNLP 2026.
- MSCE converts agent experiences into callable skills with evidence links and reliability estimates, no fine-tuning required.
- Reflection-weighted value backfilling propagates sparse feedback through self-reflections to calibrate memory evolution.
- Outperforms state-of-the-art on EvoAgentBench and LoCoMo benchmarks with strong cross-domain transferability.
Why It Matters
Enables LLM agents to learn and adapt continuously without costly retraining, crucial for real-world autonomous systems.